State of Vulnerabilities 2018/2019 - Analysis of Events in the life of Vulnerabilities

STATE OF VULNERABILITIES 2018/2019 Analysis of Events in the life of Vulnerabilities DECEMBER 2019 0

STATE OF VULNERABILITIES 2018/2019

ABOUT ENISA

The mission of the European Union Agency for Cybersecurity (ENISA) is to achieve a high common level of cybersecurity across the Union, by actively supporting Member States, Union institutions, bodies, offices and agencies in improving cybersecurity. We contribute to policy development and implementation, support capacity building and preparedness, facilitate operational cooperation at Union level, enhance the trustworthiness of ICT products, services and processes by rolling out cybersecurity certification schemes, enable knowledge sharing, research, innovation and awareness building, whilst developing cross-border communities. Our goal is to strengthen trust in the connected economy, boost resilience of the Union’s infrastructure and services and keep our society cyber secure. More information about ENISA and its work can be found at www.enisa.europa.eu.

CONTACT

For contacting the authors please use team@enisa.europa.eu. For media enquiries about this paper, please use press@enisa.europa.eu.

PROJECT MANAGER

Cosmin CIOBANU (ENISA)

AUTHORS

Vasilis Katos, Shahin Rostami, Panagiotis Bellonias, Nigel Davies, Agnieszka Kleszcz & Shamal Faily (Bournemouth University), Arnolnt Spyros, Alexandros Papanikolaou, & Christos Ilioudis (InnoSec), Konstantinos Rantos (International Hellenic University).

ACKNOWLEDGEMENTS

We wish to thank the following experts for their continuous support and guidance throught the project: Ioannis Psaroudakis (CERT-EU), James Shank & Steve Santorelli (Team-Cymru), Mihai Barbulescu (RoEdu), Michael Mortensen (Risk Based Security), Kristine Hovhannisyan & Andrew Roberts (Tallinn University of Technology), Costantinos Patsakis, (University of Piraeus), Megan Pleva (Limetools), Cagatay Yucel (Yasar University).

LEGAL NOTICE

Notice must be taken that this publication represents the views and interpretations of ENISA, unless stated otherwise. This publication should not be construed to be a legal action of ENISA or the ENISA bodies unless adopted pursuant to the Regulation (EU) No 2019/881. This publication does not necessarily represent state-of the-art and ENISA may update it from time to time.

Third-party sources are quoted as appropriate. ENISA is not responsible for the content of the external sources including external websites referenced in this publication.

This publication is intended for information purposes only. It must be accessible free of charge. Neither ENISA nor any person acting on its behalf is responsible for the use that might be made of the information contained in this publication.

COPYRIGHT NOTICE

© European Union Agency for Cybersecurity (ENISA), 2019 Reproduction is authorised provided the source is acknowledged.

STATE OF VULNERABILITIES 2018/2019

Copyright for the image on the cover is on: © Shutterstock For any use or reproduction of photos or other material that is not under the ENISA copyright, permission must be sought directly from the copyright holders. ISBN: 978-92-9204-325-4, DOI: 10.2824/139301

STATE OF VULNERABILITIES 2018/2019

TABLE OF CONTENTS

1. INTRODUCTION 7

1.1 MOTIVATION AND AIMS 7

1.2 STANDARDISATION IN VULNERABILITY MODELLING 7

1.3 VULNERABILITY MANAGEMENT 9

2. METHODOLOGY 10

2.1 RESEARCH APPROACH 10

2.1.1 Experts Group Formation 10 2.1.2 Formulation of research questions 10 2.1.3 Identification of Data Sources 11 2.1.4 Establishing a schema 13 2.1.5 Data collection, cleansing, normalisation and curation 13 2.1.6 Development of Jupyter notebooks 13 2.1.7 Analysis 14

3. ANALYSIS & FINDINGS 16

3.1 EXPLORATORY DATA ANALYSIS 16

3.2 VULNERABILITIES PER SECTOR 18

3.3 HIERARCHICAL CLUSTERING & FACTOR ANALYSIS 18

3.4 CVSS 20

3.4.1 Comparison between CVSS v2 and v3 20 3.4.2 High CVSS score vulnerabilities 22

3.5 LIFECYCLE ANALYSIS 23

3.5.1 Exploits 24 3.5.2 End of support (End of life) 25 3.5.3 Disclosure after vendor notification grace period 26 3.5.4 Publication delays 26

3.6 WEAKNESSES 26

3.7 SOFTWARE CATEGORIES 28

3.7.1 Software categories and their weaknesses 30 3.7.2 Software categories with the highest number of vulnerabilities 31

3.8 WEB 32

3.8.1 All products 32

STATE OF VULNERABILITIES 2018/2019

3.8.2 Windows applications 32

3.9 OPEN SOURCE SOFTWARE 33

3.10 POPULARITY 34

3.11 NON-CVE ANALYSIS 35

3.12 EXPLOITS, TACTICS AND PATTERNS 36

3.12.1 CAPECs 37 3.12.2 ATT&CK techniques 37

3.13 ECONOMIC ASPECTS 39

3.13.1 Bug bounty Error! Bookmark not defined. 3.13.2 Exploit price estimation 41

3.14 CONCLUDING REMARKS 42

4. REFERENCES 44 A ANNEX: LIST OF VULNERABILITY DATABASES 45

A.1 FREE ACCESS DATABASES 45

A.2 COMMERCIAL DATABASES 45

B ANNEX: JUPYTER ENVIRONMENT INFORMATION 47

B.1 README.MD 47

B.1.1 Getting started 47

B.2 LIST OF JYPTER NOTEBOOKS 48

STATE OF VULNERABILITIES 2018/2019

EXECUTIVE SUMMARY

The vulnerability ecosystem has matured considerably in the last few years. A significant amount of effort has been invested to systematically capture, curate, taxonomize and communicate the vulnerabilities in terms of severity, impact and complexity of the associated exploit or attack. Standardisation in the description of vulnerabilities contributes not only to effective threat intelligence sharing, but also potentially efficient threat management, provided that organisations, vendors and security researchers actively seek to discover the vulnerabilities and respond in a timely fashion.

As the standardisation of cataloguing and modelling the vulnerabilities reaches the aforementioned maturity, public or private (i.e. commercial) databases containing information of the actual vulnerabilities (and some with their exploits counterparts) have emerged. As there are a number of initiatives within the research community, quite naturally some databases could be considered to be more “authoritative” and/or “reliable” than others. However, due to the nature of the vulnerability ecosystem, it is not a reasonable assumption that the databases will be complete (that is, contain all vulnerabilities), or reliable in the sense that the information captured is correct, in the sense that the samples gathered can be considered to reliably help in drawing conclusions on the whole population. This is influenced by a number of factors, including the quality of analysis and assessment, the assessment framework itself, the economic aspects (such as the value of any available exploit), as well as the business models of the software vendors, threat intelligence services, and the overall security community.

The purpose of this report is to provide an insight on both the opportunities and limitations the vulnerability ecosystem offers. By using the vulnerabilities published during the year of 2018 and Q1-Q2 of 2019 as a vehicle, this report goes beyond the standard exploratory analysis, which is well captured by many industry whitepapers and reports, and attempts to answer questions related to the reliability, accuracy of the vulnerability sources and the widely accepted evaluation metrics.

In addition, the report leverages established vulnerability taxonomies and frameworks to explore and identify more intrinsic relationships and characteristics. Vulnerabilities are explored in terms of the ATT&CK taxonomy , revealing non-uniform distribution in the defined tactics:

1. Differences, inconsistencies and discrepancies between the two major versions of the scoring systems (CVSS version 2 and version 3) may influence risk management actions; 2. Vulnerabilities showing affinity to specific industry sectors, form strong clusters; and of course the 3. Position and performance of vendors and products which varies depending on the type of software.

This report is also accompanied by the underlying dataset and software developed (in Jupyter /Python). These are made publicly available to enable further and independent exploration and analysis of the vulnerability domain by the information security community as well as allow researchers to appreciate the degree of intractability surrounding empirical analysis of vulnerabilities.

STATE OF VULNERABILITIES 2018/2019

The key findings of the analysis are as follows:

 There are significant differences between the two vulnerability measurement systems (CVSS v2 and CVSS v3) regarding the underlying recorded values. This is possibly attributed to the different wording of the categorical variables fuelling subjective bias. In either case, the correlations of the three impact measures (Confidentiality, Integrity and Availability) were surprisingly low, with Integrity and Availability having a correlation coefficient less than 0.4.  There are inconsistencies and discrepancies between the different sources. Although there is an authoritative database capturing vulnerability details, this does not imply that the information in that database is accurate.  The developed taxonomies and standards to describe the vulnerabilities are indeed rich and detailed, but only a subset of the categories was present in the 2018-2019 vulnerabilities.  There are statistically significant differences between the severity level of CVE (officially recorded) and non-CVE vulnerabilities (i.e. those that were not listed or included in the CVE databases), with the latter showing a higher score.  The exploit publication date of CRITICAL vulnerabilities is attracted near the vulnerability publication date, with the most exploits being published shortly before or after the vulnerability publication date.  At least 8.65% of the vulnerabilities are exploitable. This number is expected to be higher due to zero-day exploits and the incompleteness of the datasets.  Defence Evasion, Persistence and Discovery are the preferred tactics for the exploits.  Most exploits target web and client-side related vulnerabilities.  The top 10 weaknesses account for almost two thirds (64%) of the vulnerabilities.

STATE OF VULNERABILITIES 2018/2019

1. INTRODUCTION

1.1 MOTIVATION AND AIMS

Vulnerabilities are weaknesses leveraged by adversaries to compromise the confidentiality, availability or integrity of a resource; such weaknesses may result from design choices . The market for vulnerabilities has become vibrant in recent years, with different stakeholders and threat actors taking advantage of the opportunities afforded by exploits on offer. Once publicly known, structured information about vulnerabilities is curated in public repositories such as the National Vulnerabilities Database (NVD), while unstructured information is stored and discussed on online forums and locations.

Structured information about vulnerabilities facilitates widespread and timely sharing of information. Significant efforts are now being made to standardise this information to reduce communication barriers and complexity, leading to more effective analysis of vulnerabilities and a better understanding of the context within which different vulnerabilities are discovered. These efforts are, however, fraught with difficulty due to the challenges of categorising vulnerabilities. Vulnerability data can be incomplete, inaccessible, or inaccurate, and the quality of the resulting information has an impact on decision making, policies, and practices. Moreover, the vulnerability disclosure is influenced by a variety of factors, including financial incentives, the agenda of the disclosing stakeholder, the interaction of the various actors and this is all performed in a highly dynamic information security market (ENISA, 2015, 2018).

This study aims to address these challenges by completing three objectives:

1. Represent the state of cybersecurity vulnerabilities in a form allowing stakeholders to make informed decisions on cybersecurity investments. 2. Comprehensively analyse and correlate vulnerability data to better contextualise vulnerabilities. 3. Analyse vulnerability data from a quality and reliability perspective.

1.2 STANDARDISATION IN VULNERABILITY MODELLING

Disclosed vulnerabilities are usually uniquely identified, similar to books and publications receiving an ISBN number. The most pervasive vulnerability identification and numbering scheme is the Common Vulnerabilities and Exposures (CVE) referencing system instigated by 6 7 MITRE . Identifiers are assigned by CVE Numbering Authorities (CNAs), i.e. organisations that are authorized to assign CVE IDs to vulnerabilities affecting products, and vulnerability information typically includes a brief description, advisories, mitigation measures and reports. Other numbering schemes include Microsoft’s Security Bulletin (MS), Seebug’s Vulnerability Database (SSV) and VMWare’s Security Advisory (VMSA).

As the de facto standard, other schemes now typically map to CVE entries. Although this report adopts the CVE convention, not all publicly disclosed vulnerabilities have an associated CVE- ID. Vulnerabilities kept private and not publicly disclosed are often referred to as “zero-day vulnerabilities”, and the corresponding exploits are referred to as zero-day (0day) exploits.

Certain events in the lifetime of a vulnerability and the time periods between such events can be significant in the sense that they may influence the risk of the underlying system and provide

STATE OF VULNERABILITIES 2018/2019

opportunities to the adversaries. For example, there are normally delays in recording CVE and disclosure. There are also cyclic events associated with the battle between attacker exploitation and defensive code patching. This overall chronology is referred to as the vulnerability lifecycle (see Section 1.3).

The severity of the impact of a vulnerability is defined using the Common Vulnerability Scoring System (CVSS) maintained by the Forum of Incident Response and Security Teams (FIRST). It is often provided as a qualitative value (Low, Medium or High); this is based on a quantitative calculation derived from the characteristics of individual vulnerabilities. The current CVSS version is v3.1, implemented in June 2019 replacing v3.0, but v2.0 values are often quoted for vulnerabilities prior to June 2015 when v3.0 was published.

Vulnerabilities are associated with information technology systems and software packages; these can be categorised using the Common Platform Enumeration (CPE) database . Exploits and Exploit Kits are developed to take advantage of the vulnerabilities. Exploitable vulnerabilities are usually based on software weaknesses. Vulnerabilities can also be captured using CWE™ (Common Weakness Enumeration): “a community-developed list of common software security weaknesses” and provides developers with advice on “prioritizing software weaknesses in a consistent, flexible, open manner” ; CWE entries rely on the Common Weakness Scoring System (CWSS).

A comprehensive understanding of cyber-attacks and the Cyber Kill Chain (CKC) requires awareness of the vulnerability lifecycle including development of vulnerabilities into exploits. The CKC also provides aspects of threat intelligence by assigning some attacker behaviours to specific events and uses model descriptions to comprehend those behaviours. This knowledge helps operators of targeted systems determine a successful defence strategy and solutions to certain cyber-attack problems.

The Cyber Kill Chain (CKC) (Hutchins et. al, 2010) models cyber-attacks as event sequences from reconnaissance through exploitation to command and control of defender systems to achieve attacker actions on their objectives. For example, such a chain might model an attack that also discloses confidential information via malware. By modelling threats, the CKC helps determine their severity and explains how they are enacted. The CKC was used in the ENISA annual threat landscape report 2018 (ENISA, 2019).

MITRE also maintain Common Attack Pattern Enumeration and Classification (CAPEC) which is a “dictionary of known patterns of attack employed by adversaries to exploit known weaknesses in cyber-enabled capabilities. It can be used by analysts, developers, testers, and educators to advance community understanding and enhance defences.” CAPEC provides an understanding of how adversaries operate thus supporting effective cybersecurity. This can be complemented by ATT&CK™ , which records attack tactics, techniques and procedures using matrices to map techniques to specific tactics; together this can be used to represent CKCs. The ATT&CK framework’s tactics, although not explicitly referred to as a kill chain phase, could be considered as such.

These threat intelligence standardisation activities are summarised in Figure 1.

STATE OF VULNERABILITIES 2018/2019

Figure 1: Vulnerability model context

1.3 VULNERABILITY MANAGEMENT

Early work defining the concept of the modern vulnerability lifecycle was reported by Arbaugh et al. (2000). The lifecycle is depicted in Figure 2. Mapping the vulnerability lifecycle identifies significant milestones and events that define risk transitioning boundaries. The significance of risks increases as vulnerabilities trigger the creation of the associated exploits and decrease when the patches become available.

Figure 2: Vulnerability lifecycle

STATE OF VULNERABILITIES 2018/2019

2. METHODOLOGY

2.1 RESEARCH APPROACH

The research approach is depicted in Figure 3. The main steps and phases include the formation of a panel of experts, the formulation of the research questions, the identification of the data sources, the establishment of an appropriate schema, data collection, cleansing and curation, the development of the Jupyter notebooks and the actual analysis. Although the process followed these steps in a linear manner, the involvement and intervention of the experts throughout the duration of the project introduced a degree of iterations in order to refine the quantity and quality of the research questions as well as the creation of an appropriate dataset.

Figure 3: Research approach

2.1.1 Experts Group Formation

The experts who participated and contributed to the project were from the cyber security industry and academia, as well as from an EU Organisation (CERT-EU). As this project was led by ENISA together with an academic team, it was critical to involve experts from non-academia to validate the work and establish a level of analysis and communication of the results appropriate to the target stakeholders. As such, with regards to the academic experts, Tallinn University of Technology was invited as they also maintain an academic CERT. Including experts from academic CERTs and CERT-EU enabled a better collaboration with the Industry experts.

The experts contributed in two aspects. First, they reviewed the proposed research questions in order to affirm that the research effort would address questions that would be of added value to the industry and organisations. Second, they supported the data collection process either by directing the researchers to the appropriate sources, or by contributing with their own datasets, which improved the speed of collection. It should be highlighted though that all data used in this research are open source.

2.1.2 Formulation of research questions

As the aim of this study is to allow the stakeholders to make informed decisions on cybersecurity investments through the exploration and correlation of the vulnerability data, a

STATE OF VULNERABILITIES 2018/2019

representative list of questions was established from the outset. However, since this study is accompanied with the analysis scripts and dataset, these questions as well as the attempted answers should be viewed as a starting point to empower organisations to conduct further research by enriching both the questions and the dataset.

The questions were mainly related to the evaluation of the quality and quantity of vulnerabilities, where quality denotes the information provided by vendors and third parties regarding identified characteristics of the vulnerability as well as the information related to identified exploits and the industry’s reflection.

The answers provided to these questions are anticipated to provide valuable input to the security industry with regards to the way that identified vulnerabilities are being handled, so that to improve, if necessary, their capabilities to accurate and timely identification and evaluation of exploits and remediation efficiency.

As such, issues related to the vulnerabilities and exploits sources, timings in identification, evaluation and remediation of vulnerabilities and exploits, standardisation and adoption thereof with regards to information formatting and communication will be addressed through this research.

2.1.3 Identification of Data Sources

A variety of sources were identified through the course of the study. The following types of sources were considered:

 Vulnerability Databases and taxonomies as listed in Figure 4, and articles about databases. A recent research considering both vulnerability management aspects and data sources by Kritikos et al. (2019) was consulted.  Articles on specific vulnerabilities offering more detail. For example, ThreatConnect are among several threat intelligence aggregators and providers who collect and consolidate information on vulnerabilities’ appearance.  Vendors Bulletins (essentially unstructured data)  General News Sources (essentially unstructured data)  Tools and sources for mapping vulnerability information to MITRE’s ATT&CK taxonomy.  Economics Sources (associated with black market exploit and vulnerability prices plus general market volume analysis data)

In order to compile the required datasets, these sources were studied and evaluated, to ensure that the data obtained from them would be appropriate, relevant and of sufficiently high quality. Based on the research questions that were defined in the beginning of the project, certain criteria were set, which were in turn used for performing the selection of the data sources, thus leading to the list presented in Error! Reference source not found.5. The main requirement for t he selection of each specific data source was that the data it provides must be free of charge and primarily in the form of structured datasets. The information provided by a data source was also evaluated in terms of accuracy, consistency and completeness by considering additional external references from other well-established sources or standards (e.g. CVE, BID, CWE, etc.). The ease of data extraction from a given source was an additional quality that was taken into consideration.

The selected data sources were divided into three categories: a) The ones that had a CVE ID assigned to the vulnerability data they provided (CVE data), b) the ones that did not (non-CVE data) and c) the ones that provided information relating to CVEs (CVE-related), such as CAPEC IDs, CWEs, etc.

STATE OF VULNERABILITIES 2018/2019

Although locating CVE data and CVE-related sources was relatively easy, the collection of non- CVE data proved to be a more difficult task, primarily because such sources are limited in number and they normally provide their data for a fee. Nevertheless, non-CVE vulnerabilities may represent a significant proportion of all vulnerabilities.

Figure 4: Data sources

*Commercial database

The data sources were combined in order to produce a rich dataset with a variety of features (or dimensions). As many of the sources had feature overlaps it was possible to create a superset contextualising each vulnerability further. For instance, the CVE ID was used to link information between the weaknesses, industries, number of exploits and incidents. The CAPEC that was included in both NVD and ATT&CK allowed mapping of a CVE to different attack techniques and tactics.

STATE OF VULNERABILITIES 2018/2019

Combining different sources to produce one dataset inevitably led to having empty values, as the different data sources do not necessarily overlap horizontally. However, the number of vulnerabilities is adequate (over 27K) to reliably conduct statistical tests and in most cases the sample sizes where adequate to draw safe conclusions.

2.1.4 Establishing a schema

The data schema was developed on the basis that the analysis would be conducted through a collection of purpose-built Python Jupyter notebooks. As such, priority was given to having a dataset in a form that is most suitable for a sequential analysis utilising the notebook(s). Moreover, it was possible to estimate the size of the dataset reasonably accurately, both in terms of features/columns as well as number of observations/elements/rows. Furthermore, it was recognised that flat files (that can be easily imported into Python’s Pandas data frames) offer advantages over maintaining other and potentially more complex structures such as hierarchical databases.

2.1.5 Data collection, cleansing, normalisation and curation

This study covers the period of vulnerabilities published between January 1 2018 to August

31 (Q1 – Q3) 2019. The vulnerabilities were regularly collected and hosted in the compiled

dataset until the cut-off date of September the 30 . As such, the NVD “snapshot” of the dataset reflects that date and there are likely to be more vulnerabilities included in the period under examination today. This is due to the lag of vulnerabilities getting officially a CVE ID and entering the system.

In order to further contextualise the vulnerability entries, several dictionaries were also downloaded, see Figure 5: These files were used through mapping and lookup functions to translate the numerical values and IDs to more meaningful information.

Figure 5: Dictionaries

It should be noted that some of the dictionaries did not include complete information leading to further missing values. For instance, the CPE list did not include a complete mapping between the CPE vector and product/vendor information. Although some entries were manually created and added, it is highlighted that the incomplete information is common among such data sources.

2.1.6 Development of Jupyter notebooks

The analyses have been conducted within Jupyter notebooks. The open-source Jupyter Notebook technology enables the creation of documents using the literature programming paradigm, where code, narrative, and interactive visualisations can be blended together. These

STATE OF VULNERABILITIES 2018/2019

documents can be easily shared or downloaded from the web, and contain live code which enables complete reproducibility.

2.1.7 Analysis

The analysis employed - where appropriate - statistical methods and techniques to assess the significance of the findings. Although this report that accompanies the Jupyter notebooks contains fewer details on the analysis, it does however summarise the main and most significant findings or those that the contributors deemed to have the highest potential, added value and impact. Moreover, the statistical approaches presented in the report are explained in a higher level of detail in the Jupyter notebooks.

2.1.7.1 Statistical tools

A basic test for any distribution is to check whether the distribution is normal, as in such case it would enable one to run a wide variety of statistical tests, namely parametric tests. A caveat however when running normality tests on large (>100) samples, is that most tests fail as they are sensitive (strictly speaking, they are dependent on the standard error which becomes very small as it includes the square of sample size in its denominator). As such, it is also important to visually inspect the distribution and decide whether parametric tests could still be used, accepting of course lower confidence trade-offs. This work employs Kolmogorov-Smirnov’s test for normality.

Distribution comparison refers to checking whether the means of two distributions (or samples) are equal. Wilcoxon test is the most common test and used in this report.

The independent t-test is a popular approach to test the hypothesis of two samples having (significant) different means. The t-test is applicable to variables that have two levels (e.g. before and after, or present/absent). If the variables have more than two levels, analysis of variance (ANOVA) is performed. One-way ANOVA is performed to compare the means of three or more groups of data over one independent variable, whereas two-way ANOVA is used to compare the means of three or more groups over two independent variables. An example of a one-way ANOVA is to check if the CVSS means of the ATT&CK tactics (groups) are different. Two-way ANOVA can be used to check for instance if there is significance in the interaction between publication delays and severity of a vulnerability.

Both hierarchical clustering and factor analysis aim to explore more intrinsic relations. Hierarchical clustering is performed by considering each data sample as a vector and measures the distance from the other data items. The output can be in the form of a dendrogram, where the leaves represent the factors or variables and the length of the paths between them show their distance. The dendrogram is structured in a way that the more we travel towards the root, the bigger the clusters that are formed, reducing thus the total number of clusters. Normally, a horizontal line threshold is drawn about halfway through the height of the dendrogram and the number of intersecting branches show the number of clusters. In this work Ward’s approach is used.

While hierarchical clustering gives a high level and visual representation of clusters, factor analysis can provide a more detailed and quantitative description of the potential factors. Factor analysis aims to consolidate and group variables with a view to perform dimension reduction over the data variables. A necessary condition to perform factor analysis is first to check if the correlation matrix is not the identity matrix. In this work, this is performed by Bartlett’s sphericity test.

Following factor analysis, Chronbach’s alpha measures the internal consistency of variables (items) and indicates how closely related they are as a group. This measure is used to assess whether the variables can be combined to indeed form a factor (also referred to as a latent, unobserved variable). Normally for values over 0.70 it is commonly accepted that the variables

STATE OF VULNERABILITIES 2018/2019

can indeed be grouped. For instance, the CVSS formula groups Confidentiality, Integrity and Availability into the Impact subcategory. The Cronbach’s alpha value for these three features is 0.83 which is rather high and shows that the choice to have these three under one group was a correct design decision. Moreover, this high value is particularly interesting as it directly relates to the main goals and essence of information security.

2.1.7.2 Economic aspects

Analysis of the economic aspects of vulnerabilities is a critical, complex, and sometimes more esoteric to the topics discussed above. The economic and financial impact is closely coupled to an organisation’s risk assessment and risk management approach and is outside the scope of this study. However, this report considers the economic aspects of the development of exploits and includes a high-level analysis of the exploit prices. To this end, bug bounty prices were included (from Zerodium) as well as the exploit price estimates by VulDB who maintain a proprietary estimation algorithm.

In the next Section all findings presented are of statistical significance, when an applicable test was employed. To conduct further and independent data exploration refer to the Jupyter notebooks collection at https://github.com/enisaeu/vuln-report.

STATE OF VULNERABILITIES 2018/2019

3. ANALYSIS & FINDINGS

3.1 EXPLORATORY DATA ANALYSIS

As mentioned earlier, the vulnerability dataset contains missing values due to the missing data from the source datasets but also due to the joining operation of the different sources. Figure 6 summarises the sample size of the respective columns.

Figure 6: Main dataset items population

Word cloud output from the main dataset’s description field

As mentioned in the previous section, the main source of CVEs is the National Vulnerability Database. As many organisations use this as the authoritative source of CVE entries, it is worth investigating the completeness of this source. It was found that some CVEs that appeared in other databases, did not have up to date entries in NVD. Consider for example, CVE-2018-

STATE OF VULNERABILITIES 2018/2019

14319 (CVSS score 6.8) which is a buffer overflow / Remote Code Execution vulnerability affecting Samsung Galaxy S8 smartphones. This vulnerability:

 was reported to the vendor in Q1 of 2018, (April 2018, ZDI ),  was reported by the vendor through an initial advisory in August 2018 (SVE-2018-11828), classifying it as a CRITICAL level vulnerability ,  did not secure an entry in NVD .

From the above the following observations are made. First, organisations that rely solely on one source – no matter how authoritative it may be – will potentially miss vital vulnerability information affecting their systems. Second, the CVSS scoring system, although capable of providing potentially a good reference for assessing, understanding and contrasting the impact and severity of vulnerabilities, does not necessarily forms the “ground truth” as other stakeholders may have a different view on the severity such as Samsung in this case who are the vendor of the vulnerable product.

A non-uniform distribution of vulnerabilities over the ATT&CK tactics is evident in Figure 7. Defense Evasion is clearly by far the most “popular” tactic for vulnerabilities, whereas three tactics do not have any representation in the 2018 dataset. It should be noted that some vulnerabilities are counted more than once, if they appear in multiple tactics.

Figure 7: Distribution of vulnerabilities over ATT&CK tactics and CVSS v3 base score (29.37% of vulnerabilities in dataset)

The ATT&CK framework is constantly enriched with techniques and sub-techniques. At the time of writing, the number of techniques recorded to 291. In the 2018 dataset, 52 techniques were associated with vulnerabilities, accounting to the 17% of the complete ATT&CK techniques range. The top 3 techniques associated with vulnerabilities are:

T1148 – HISTCONTROL (4226 vulnerabilities) T1027 – Obfuscated Files or Information (2293 vulnerabilities) T1130 – Install Root Certificate (1813 vulnerabilities)

STATE OF VULNERABILITIES 2018/2019

Out of the 52 techniques, 35 have over 1600 appearances in vulnerabilities, following with a large dip (443 vulnerabilities and below) for the remaining 17. The techniques are explored further in Section 3.12.

3.2 VULNERABILITIES PER SECTOR Vulnerabilities

Figure 8 summarises the vulnerabilities per sector as well as the number of critical

with the highest

vulnerabilities for each sector.

cross-sectoral

Figure 8: Vulnerabilities per sector impact:

CVE-2019-1639 CVE-2019-1638 CVE-2019-1636 3.3 HIERARCHICAL CLUSTERING & FACTOR ANALYSIS

Clustering and factor analysis can show more intrinsic relationships existing in the data. Starting with hierarchical clustering, Figure 9 shows the dendrogram after performing Ward’s distance approach on the ATT&CK tactics.

Figure 9: Hierarchical clustering results on ATT&CK tactics

STATE OF VULNERABILITIES 2018/2019

From the above there are some clear clusters emerging: Collection & Command and Control, which then can be grouped with Credential Access, and further with Discovery. On the other side, Privilege Escallation with Persistence form a cohesive cluster which join Execution and Lateral Movement.

However, this approach provides a more qualitative and intuitive representation. A more quantitative and detailed approach is achieve through factor analysis. Following this approach, the results are summarised in Figure 10. It should be noted that the initial analysis suggested 3 factors, but following an internal consistency check (using Cronbach’s alpha), 4 factors were confirmed. The loadings column shows the weight of each of tactic in the respective factor.

Figure 10: Factor analysis results.

STATE OF VULNERABILITIES 2018/2019

Figure 11: Hierarchical clustering results on Sector (Zero Day Initiative sample)

A similar approach was followed for the sector categories, Figure 11. Quite interestingly,

Vulnerabilities

Transportation Systems are closely coupled with Chemical and are distant from Transportation

in Energy and

Services. A clear cluster is shown for Water, Food and Agriculture, Chemical, Transportation and Commercial Facilities. Vulnerabilities in Energy are quite distinct from those found in the Critical other sectors. Manufacturing

are distant 3.4 CVSS from all others

CVSS scores are important metrics as they provide a quantitative measure that can be eventually used to inform risk exposure. Organisations use CVSS to make judgements on their vulnerabilities based on this metric. In the following analysis the CVSS versions are compared and contrasted to establish whether the different measuring systems have differences leading to potentially different decisions.

3.4.1 Comparison between CVSS v2 and v3

Figure 12: Distribution of CVSS v2 and v3 base scores

STATE OF VULNERABILITIES 2018/2019

It is visibly evident from Figure 12 that the distributions using v2

and v3 are significantly different, with CVSS v3 having a considerably higher mean (7.34) than CVSS v2 (5.75). The

biggest difference was observed for CVE-2019-12373 (Ivanti

LANDESK Management Suite) with a value of 6.3 (CVSS v2: 2.7, CVSS v3: 9.0)

Severity and impact are contrasted in Figure 13. Again, there are visible differences between

There are 4

the two CVSS versions. In this case, the severity for v3 was higher than that of v2, but the

potentially

opposite was true for impact (CVSS v2 higher impact than v3). Furthermore, the correlation results for the three impact categories were as follows: misclassified

CRITICAL and

 Confidentiality: 0.79 exploitable

 Integrity: 0.34

vulnerabilities

 Availability: 0.38

affecting confidentiality,

The correlations could arguably be considered low, given that both versions have three levels per impact variable (Confidentiality, Integrity, Availability), but different qualitative definitions, 42 for integrity,

namely for version 2: “none”, “partial”, “complete” and for version 3: “none”, “low”, “high”. The 46 for availability

different definitions could be sufficient to be responsible for the associated bias when assigning

the values.

Figure 13: Severity and Impact

Exploitability attempts to capture the ease (or complexity) by which the vulnerability can be

exploited. Although the underpinning metrics slightly differ between versions 2 and 3,

CVSS V2 & V3

exploitability shows the highest contrast between the two CVSS versions (Figure 14). It should There are be noted that these two versions have different definitions leading to different ranges with considerable version 3 defined in a range of 0-3.9, whereas version 2 operates in the range of 0-10 and over differences between three quarters of vulnerabilities have an exploitability score over 8. the two scoring

systems leading to The correlation coefficient between the two exploitability versions is 0.79, which could be different severity considered marginally acceptable, given that both versions represent the same aspect. classifications, that Inspecting the different components of exploitability, those that showed the highest correlation may in turn affect the (between the two versions) were user interaction (v3) with access complexity (v2) with a risk management, correlation of 0.85, and attack complexity (v3) with access complexity (v2), with a value of planning and 0.78. decision-making

processes.

STATE OF VULNERABILITIES 2018/2019

Figure 14: Exploitability

3.4.2 High CVSS score vulnerabilities

The following series of graphs refer to vulnerabilities with score equal or greater than 7 (High, Critical). Figure 15 shows the top 20 products with the most (and high) CVEs and Figure 16 the top 20 vendors with the most CVEs. Figure 17 presents the mean scores of the vulnerabilities by vendor and ATT&CK tactic.

Figure 15: Top 20 products with the most CVEs (having score >=7)

The product with the highest number of vulnerabilities is the Android OS, followed by Debian Linux and Acrobat reader. The list is dominated by Qualcomm’s Sd series Firmware. Further down the list lie more Windows based operating systems whereas Edge is in position 26. Note that this distribution considers vulnerabilities by products; that is, if a particularly vulnerability affects more than one product it will be counted twice (or more times for that matter, depending on the number of products it affects). A different view of the above is presented if the vulnerabilities are grouped by vendor, where in this case Qualcomm is a runner-up to Microsoft (Figure 16).

STATE OF VULNERABILITIES 2018/2019

Figure 16: Top 20 vendors with the most CVEs (having score >=7)

Microsoft has the highest number of vulnerabilities (600) which is more than 50% higher than the runner-up, Qualcomm. From an ATT&CK framework perspective (Figure 17), Cisco and Canonical have high scored vulnerabilities across the range of the tactics, whereas Microsoft is ranked among the lowest.

Figure 17: Vendor vs. ATT&CK tactic vs. CVSS score (mean) heatmap

3.5 LIFECYCLE ANALYSIS

Among the key milestone dates of a vulnerability is its actual publication date by which the vulnerability becomes widely known, and the publication of the exploit; both these milestones normally put pressure to the vendor to come up with a security update. In a relevant note, End of Support (EOS) date is the point in time where the vendor is not expected to develop a patch. These are explored in the following sections.

STATE OF VULNERABILITIES 2018/2019 3.5.1 Exploits

Figure 18 shows the percentage of vulnerabilities with exploits and their publication referencing EXPLOIT the vulnerability publication date (t=0).

PUBLICATION DATE

Figure 18: Exploit publication date (t=0: vulnerability publication date, sample size 8.63% of Exploits of dataset) vulnerabilities of CRITICAL severity level form a constellation around the vulnerability publication date.

Note that the above results are expected to be skewed due to the upper-bound / cut-off time of the data. In order to proceed with the analysis, we need to ensure that the data are not biased. As such, we select only the vulnerabilities published in 2018 and allow a 6-month window (180 days) for the exploits. The distribution and descriptive statistics are presented in Figure 19. The descriptive statistics show the mean of difference per severity.

Figure 19: Exploits for 2018 vulnerabilities published within 6 months (before or after) of the publication date of the vulnerability.

Note that for CRITICAL severity levels, the mean is smaller than all other severity levels (24.83). A 2-way analysis of variance (ANOVA) shows that there is indeed significant difference between CRITICAL and all other severity levels, with the former clustered around the vulnerability publication date.

STATE OF VULNERABILITIES 2018/2019

Figure 20 shows the outlier products with the highest (and fewest) exploits published before or after the publication of a vulnerability. The results suggest operating systems attract more exploits after the publication of a vulnerability.

Figure 20: Publication of exploits surrounding vulnerability publication date.

3.5.2 End of support (End of life)

A pivotal moment in the life of a software product is when the vendor ceases to support it. From a security perspective this is critical as it stops issuing patch updates. One of the reasons malwares like WannaCry created havoc in a short timeframe was because it exploited vulnerabilities of systems that were impossible to patch, simply because the vendor stopped supporting them.

Figure 21 shows the distribution of exploits published before or after the End of Support (EoS) date. The two groups (before and after EoS exploits) do not have equal variances, but there are no significant differences between their means.

Figure 21: End of support exploits.

STATE OF VULNERABILITIES 2018/2019 3.5.3 Disclosure after vendor notification grace period

When vulnerabilities are reported to the vendor, it is expected that they would issue a patch within a certain timeframe. This timeframe should be subject to the severity of the vulnerability, however this is not observed. Zero Day Initiative (ZDI) for example have a 120-day grace period. Following that, they disclose the vulnerability to the public, if the vendor does not issue any updates. The ZDI dataset contains 6 vulnerabilities (out of 2228) that fall into this category, 3 of which have a CVSS v3 score of 10 and only 2 have a CVE ID, whereas the other 4 are non-CVE (0-days). The high non-CVE vulnerability refers to Belkin’s Wemo Link (ZDI-CAN- 5206).

3.5.4 Publication delays The top 10 CWEs

The following results refer to the ZDI dataset. On average, a vulnerability is published after 112

account for the

days after it has been reported (Figure 22). Also, it takes longer for CRITICAL and HIGH

64.20% of the

vulnerabilities to be published, than those with a LOW severity score.

vulnerabilities

Figure 22: Publication delays for vulnerabilities

Operating ANOVA results: systems are at the top of unique CWEs, with the greatest variety of weaknesses 3.6 WEAKNESSES

Weaknesses are the causes leading to vulnerabilities. In the following set of graphs, the CWE data are analysed and presented.

Figure 23 presents the top 10 CWEs. Note that CWE with ID 264 refers to a category rather than a specific weakness, namely “Permissions, Privileges and Access Controls”.

STATE OF VULNERABILITIES 2018/2019

Figure 24: Average CVSS score for top 10 weaknesses

Figure 25 shows weaknesses per product, both in total and unique. Interestingly, Operating Systems dominate both lists, with Open Source OSs leading the charts. Figure 25: CWEs per product

STATE OF VULNERABILITIES 2018/2019

The corresponding chart with the vendors having the highest variety of vulnerabilities is shown in Figure 26. Debian, Redhat, Canonical and Cisco have over 55 distinct vulnerabilities.

Figure 26: Top 10 vendors with unique CWEs

Injections and improper input validation dominate the top 10 CWEs with a high CVSS score (Figure 27). It is reminded that ID 264 refers to a category rather than a specific weakness, namely “Permissions, Privileges and Access Controls”.

Figure 27: Top 10 CWEs with high CVSS score

3.7 SOFTWARE CATEGORIES

The following results refer to the 12.26% of the dataset, to vulnerabilities accompanied with a software category label, according to Vuldb. Although this subset failed the comparison test (that is, the distribution of vulnerabilities with description is different from the distribution of those lacking a description), it can be seen from Figure 28 that the two distributions are macroscopically similar and as such the subsequent analysis can be generalised for the whole vulnerabilities dataset.

STATE OF VULNERABILITIES 2018/2019

Figure 28: Comparison of vulnerabilities with and without software category description

count: 24,102 3,369 mean: 7.374 7.089 st.dev.: 1.612 1.598

Figure 29 summarises the share of the categories that have more than 10 critical vulnerabilities and those that have at least one exploit (of any severity level). In both cases the Web browser category is on the top; it should be noted though that there are no exploitable vulnerabilities of critical severity (CVSS v3 score greater or equal to 9). The total number of software categories is 54.

Figure 29: Distribution of application types for more than 10 vulnerabilities and at least one exploit (of any severity)

21 software categories having more than 50 vulnerabilities are ranked over their mean CVSS v3 base score in Figure 36. The ranking was in accordance to ANOVA test, showing the statistically significant differences between the means. Although web browsers have the highest number of vulnerabilities as shown in the previous figure, log management, multimedia player and SCADA software have a higher average severity (base) score.

STATE OF VULNERABILITIES 2018/2019

Figure 30: Top categories with high CVSS v3 base score (having more than 50 vulnerabilities).

3.7.1 Software categories and their weaknesses CWE 119:

Figure 31 tessellates the most popular software categories and weaknesses pairs. The highest Improper

number of occurrences of weakness is Improper Restriction of Operations within the

Restriction of

Bounds of a Memory Buffer which is found on 165 vulnerabilities affecting Web Browsers.

Operations within the

Figure 31: Top software categories – weaknesses pairs.

Bounds of a Memory Buffer, io is a common li ns, e weakness in Web s c ds tri Browser c rm on Mem software STATE OF VULNERABILITIES 2018/2019 3.7.2 Software categories with the highest number of vulnerabilities

Figure 32 shows the top 10 software categories with the most vulnerabilities. Out of these, the top 3 are further examined to establish if there are statistical differences in their means. Following the Mann-Whitney test, Web Browsers (mean base score: 7.39) have a significantly different (higher) CVSS v3 base score than the two runner ups, Operating Systems (mean score: 6.9) and Content Management Systems (mean score: 6.87).

Figure 32: Top 10 categories (highest number of vulnerabilities)

Figure 33 shows the top 20 software categories having most vulnerabilities, against the ATT&CK tactics. Note that as many points had a low number of observations, the maximum values are displayed.

Figure 33: Top 20 software categories against ATT&CK tactics (max base score values)

STATE OF VULNERABILITIES 2018/2019 3.8 WEB 3.8.1 All products

The following results refer to all products that relate somehow to web applications or services, relating to the 7.6% of the data. The selection process involved the inclusion of those vulnerabilities containing the keyword “web” in their description. The corresponding exploits are shown in Figure 34. Webapps type of exploit and PHP are the most popular focus of exploits.

Figure 34: Target platforms and exploit types

With regards to vulnerabilities on web products Cisco displays by far the most vulnerabilities (over 250), whereas the follow up vendor, Microsoft, has less than 100. The variation of vulnerabilities within the top 10 most vulnerable web products is limited, approximately between 37 to 57 (Figure 35).

Figure 35: Vendors and products with the most vulnerable web products

3.8.2 Windows applications

The CVSS scores and differences between CVSS v2 and v3 in Windows applications follows the overall distributions as presented earlier, with v3 showing a higher mean base score. Unsurprisingly, Microsoft as a vendor and its products dominate the top 10 vulnerabilities list (Figure 36).

STATE OF VULNERABILITIES 2018/2019

Figure 36: Vendors and products with the most vulnerable Windows applications

Figure 37 shows the severity and base scores of web and windows applications. As the CVSS base score distributions are clearly distant from a normal distribution, the Mann Whitney non parametric test was performed in order to check if these two sets differ. Indeed, the probability of the test was virtually equal to zero. Hence, we accept the hypothesis of the two samples having different means and in this case the Windows applications have a higher mean CVSS base score than web applications (7.25 as opposed to 7.1).

Figure 37: Severity ratings for web and windows applications

3.9 OPEN SOURCE SOFTWARE

The subset containing the vulnerabilities of the open source was extracted from the main dataset manually, after inspecting the rows and isolating the most popular open source projects. This led to a dataset of 3,221 vulnerabilities which accounts for the 11% of the whole dataset. The CVSS score distribution and differences between version 2 and version 3 of the sample followed the complete dataset (Figure 38). Once more, the discrepancy between the two versions is considerable.

Figure 38: CVSS v2 vs. v3 in open source software

Figure 39 shows the vendors and the projects of the open source software where Mozilla and Firefox are in the lead.

STATE OF VULNERABILITIES 2018/2019

Figure 39: Distribution of vulnerabilities on vendors and projects of Open Source software

In Figure 40 the top weaknesses in open source software are presented. CWE 255 refers to a weakness category, Credentials Management.

Figure 40: CWEs in open source software

3.10 POPULARITY

Popularity refers to the number of times a particular vulnerability is referenced by articles, posts, whitepapers, etc. In the following charts the data presented refer to the number of appearances (or “incidents”) as these have been summarised and collected by ThreatConnect.

Figure 41 summarises the 20 most popular vulnerabilities. The colour code maps to the CVSS v3 score. From a visual inspection there is no correlation between the popularity of a vulnerability and the CVSS score (to be precise the correlation is very low, 0.084). Nevertheless, the most “popular” CVEs with the highest score (CVE-2018-4878 and CVE-2019- 0708) refer to Redhat, Microsoft, Adobe and more particularly their Operating Systems software and flash player, in the case of Adobe. CVE-2018-8174 which is the vulnerability with the highest number of appearances refers to Microsoft’s Windows Server products.

STATE OF VULNERABILITIES 2018/2019

Figure 41: “Celebrity” vulnerabilities

3.11 NON-CVE ANALYSIS

A considerable amount of activity surrounds vulnerabilities that do not enter the CVE ecosystem or if they do so, it happens at a very late stage. The main dataset for example, contained vulnerabilities that received a CVE ID even 5 years after an exploit was published. In this section the Zero Day Initiative (ZDI) data were analysed. This dataset was selected as it contained structured information and an adequate number of non-CVE vulnerabilities that were scored based on the CVSS v3 system. It should also be highlighted that the data that considered to be within the scope/time range of this study are those that ZDI used an ID of the form ZDI-18-xxx. From the findings shown in Figure 42, the non-CVE vulnerabilities have a statistically significant different (higher) CVSS v3 score mean from those with CVE.

In terms of percentage between the CVE and non-CVE vulnerabilities, for the ZDI data the non- CVE vulnerabilities account for slightly over 8%, whereas VulDB reports this figure to be up to approximately 30%. This is because ZDI is primarily a subset of VulDB who gathers vulnerability information from a wider number of sources.

Figure 42: CVE vs non-CVE in ZDI data

STATE OF VULNERABILITIES 2018/2019 3.12 EXPLOITS, TACTICS AND PATTERNS

A threat actor can only succeed in an attack if they manage to develop and deploy an exploit

LOWER

against an existing vulnerability. The existence of exploits and their characteristics (such as

BOUND OF

complexity, privileges required, and so forth) can significantly affect the level of risk. In this

NUMBER OF

section the exploitation aspects of the vulnerability ecosystem are explored. According to the dataset, the lower bound of the percentage of vulnerabilities being exploited in the wild is EXPLOITS 8.65%. At least 8.65% of the vulnerabilities can be Out of the recorded 2,371 exploitable vulnerabilities, 492 have a CRITICAL (version 3) severity practically exploited. score. The top 10 are presented in Figure 43. Out of these 492 vulnerabilities, 4 had 10 or more published exploits (CVE-2015-2003, CVE-2014-2048, CVE-2015-2000, CVE-2015-2001), all This accounts to having CVSS score 9.8. 2,377 exploitable vulnerabilities for There is a 0.46 (positive) correlation between the number of exploits and the popularity (number 2018 (and half of of incidents) of a vulnerability. CVE-2018-4878 and CVE-2019-0708, both having a v3 score 2019) equal to 9.8, have a staggering number of 45 and 44 incidents published by ThreatConnect.

Figure 43: Top 10 most critical exploitable vulnerabilities

In total there are 11 platforms that have more than 20 exploitable vulnerabilities (Figure 44). PHP has the highest number of vulnerabilities, followed by Linux and Windows.

Figure 44: Platforms with more than 20 exploitable vulnerabilities

STATE OF VULNERABILITIES 2018/2019 3.12.1 CAPECs

Figure 45 summarises the most frequent CAPECs for critical vulnerabilities (score>9) with an available exploit. The 4 most frequent CAPECs exploiting over 90 vulnerabilities refer to web and client-side exploits (note that Cross Site Identification is essentially a form of Cross Site Request Forgery that does not require user actions).

Figure 45: Top 20 most frequent CAPECs for exploitable critical vulnerabilities

Remote, web

An analysis of variance carried out across all 2,377 exploitable vulnerabilities, indicated that the CVSS v3 score for remote attacks is the highest, followed by local and webapps (which have based attacks statistically the same score), followed by DoS (Figure 46). It is noteworthy that similar results targeting web were obtained with CVSS v2 base score, but with significantly lower means (ranging from 5.44

clients are the

to 6.5).

most common attack pattern

Figure 46: ANOVA results for exploit type CVSS v3 base score differences

3.12.2 ATT&CK techniques

There were 132 identified exploitable critical CVEs with ATT&CK technique identifiers. Figure 47 shows the distribution of techniques against the matching vulnerabilities. 26 vulnerabilities were

STATE OF VULNERABILITIES 2018/2019

shown to be open to as many as 19 ATT&CK techniques. Out of these, HISTCONTROL was the most popular technique, affecting 49 critical vulnerabilities, followed by Obfuscated Files or Information which is applicable to 35 vulnerabilities. The arrangement of critical vulnerabilities on the ATT&CK framework is shown in Figure 48. Persistence and Discovery are the most popular tactics, followed by Defense Evasion and Privilege Escalation.

Figure 47: Number of techniques affecting critical exploitable vulnerabilities

Figure 48: Critical exploitable vulnerabilities positioned in the ATT&CK framework – The number in parentheses represents the number of vulnerabilities affected by the technique.

STATE OF VULNERABILITIES 2018/2019 3.13 ECONOMIC ASPECTS

Economic analysis of vulnerabilities comes with a different set of challenges. The technical description of a vulnerability is expected to be fairly objective; CVE-2018-1163 for example refers to Quest’s Netvault backup product and is a critical vulnerability as it fully affects all three impact categories (confidentiality, integrity and availability), and since this can be performed with a low complexity, and no privileges required it yields a v2 base score of 10 and a v3 base score of 9.8.

STATE OF VULNERABILITIES 2018/2019

However, when studying the economic aspects of vulnerabilities, there is higher inherent subjectivity across two directions. First, there is the more esoteric direction of the impact of a particular vulnerability to an organisation. This impact feeds into the risk assessment process. The CVSS scoring system caters for refinements of the base score by introducing temporal and environmental metric groups. The latter in particular is an attempt to take into consideration the particular user’s environment in order to further contextualise the vulnerability. Naturally, this metric is rarely published in a publicly available vulnerability database. As such, this aspect of economic analysis was excluded from this study.

Second, the economic aspects of vulnerabilities can be indirectly assessed through the prices of the associated exploits. A number of sources and initiatives were analysed and it is conjectured that a considerable amount of unsubstantiated evidence exists in the wild – or to be more precise, it was not always possible to assess the validity and correctness of the claims relating to the publicly available economic data. The analysis that follows considers the price estimates from Vuldb who use a proprietary algorithm as well as Zerodium’s bug bounty programme. It can be evident from the findings below that there can be great ranges and discrepancies in the price of an exploit, which could be compared to the complex dynamics of a market led by opportunity costs.

3.13.1 Bug bounty HIGHEST BUG

Figure 49 presents the maximum awards for exploits for the major systems as advertised by

BOUNTY

Zerodium. The mobile operating systems are in the lead with Android having the higher maximum (outlier) value set to 2.5 million dollars. Although Google has paid since 2010 over PAYOUTS $15m in total to researchers to date , the highest amount recorded in their bug bounty Zero-click, zero-day programme is for bugs discovered for the Titan M chip used in Pixel smartphones, which may exploits for Android 21 OS are the highest reach a maximum of $1.5m . paid exploits In addition to the differentiation of prices based on the different systems, the bug bounty according to scheme notes high payouts for exploits delivering Local Privilege Escalation (LPE) and if done Zerodium’s bug through Remote Code Execution. The highest payout which is for Android systems refers to a bounty programme zero-click exploit, that is a 0day exploit that takes over an Android phone with no interaction from a user.

Figure 49: Zerodium’s bug bounty upper limit prices for the major systems (in $)

STATE OF VULNERABILITIES 2018/2019 3.13.2 Exploit price estimation

The analysis was performed on the publicly available subset from VulnDB. Following the ANOVA results on the CVSS base score over the four price categories (0-day low, 0-day upper, today low, today upper) it seemed that, as expected, the price is dependent on the score, with the 0-day values showing more significant correlations. The price is distinctively highest in particular for vulnerabilities of critical severity (Figure 50).

Figure 50: VulDB’s exploit price estimates

Figure 51 shows the exploit prices over the base score and in relation to the existence of a remedy (official fix, workaround, not defined). Although there are no noteworthy differences and patterns internally in the graphs, “Today’s price” values seem to be close to a “right shift” in the price band from the 0-day price values.

Figure 51: Exploit prices based on existence of remedies (fixes) and state (0-day vs. current/today’s price)

Figure 52 shows the dendrogram following hierarchical clustering on the prices; the upper price data for 0-day was excluded as it included an open upper bound value and would therefore yield less realistic results. Moreover, only the software categories with more than 40 observations were included. According to the clustering output, Web Browsers are clearly distant from all other categories. Setting a threshold at $20k, there are three clusters in total, as seen by the different colouring.

STATE OF VULNERABILITIES 2018/2019

Figure 52: Hierarchical clustering of software categories on Today’s prices and 0-day low prices (for categories with >40 recorded vulnerabilities)

3.14 CONCLUDING REMARKS

The analyses presented in the previous sections are an indication of the potential to generate intelligence, make informed decisions, and perform risk assessment exercises on software vulnerabilities. The systematic efforts of the computer security community to create a taxonomy, and develop databases populated by structured vulnerability descriptors, paves the way to a deeper exploration of the vulnerability ecosystem.

The results and findings show potentially revealing relationships and crucial information on how vulnerabilities are spawned from weaknesses. These fuel the development of exploits and the formation of attack patterns, techniques, and tactics. At the same time, when assessing the quality and accuracy of the collected data, we caution the reader to be aware of two main caveats. Firstly, the frameworks and metrics developed to express and assess the vulnerabilities are in a journey of their own: they go through continuous transformation, revision, and development. This suggests that future, more improved and rigorously validated frameworks may not be “compatible” or in agreement with the current/preceding versions, affecting the interpretation of the outcomes. Moreover, the taxonomies, frameworks, and metrics have grown to become exceedingly elaborate and detailed; whereby the approximate sample size of 27k observed vulnerabilities over the study period did not cover the whole range of categories. Although this is not necessarily a drawback for instruments such as the ATT&CK framework, it may be an issue for the CVSS metric groups, as there are substantial differences between the two prevailing quantitative descriptor versions (CVSS version 2 and version 3.x). For instance, the two CVSS versions have different definitions for the exploitability metric with significantly different ranges (where the version 2 range is 0-10 and the version 3 range is 0- 3.9). Therefore, it is critical to be clear on the version of the CVSS metrics employed when conducting any analyses, visualisations, or reports. In addition, a possible migration from one version to the other must be carefully planned and executed.

Secondly, the datasets found in both “authoritative” sources and in the wild show significant discrepancies which in some cases can cause a substantial (mis)classification of a vulnerability in terms of its severity and impact. Moreover, due to the nature of this problem domain, the vulnerability datasets are expected to be incomplete, however, we do not consider this to be a critical issue in contrast to the discrepancies caveat. We make this assumption based on the observation that fewer samples and data points mostly follow the statistical behaviour of the

STATE OF VULNERABILITIES 2018/2019

greater population, with the exception of non-CVE vulnerabilities that were found to differ from those registered under the CVE scheme.

When considering CVSS scores in particular, there are differences both in the definitions between the two scoring versions as mentioned above, as well as in the actual values themselves. The latter discrepancies may affect the reliability and trustworthiness of these quantitative schemes. For instance, although the overall impact metric formula is different between versions 2 and 3, with the latter version having the impact dependent on the Scope variable, the main impact factors used in the calculation (Confidentiality, Integrity, Availability) are not consistent within the same database, yielding alarmingly low correlations. As a result, when using vulnerability data, decision makers and risk assessors should either consider the worst-case scenario by accepting the higher impact value or, if resources permit, they should independently assess the impact of the underlying vulnerability. In any case, the risk assessment process should integrate outlier and discrepancy detection layers and functions into the underlying risk assessment toolbox.

This publicly available report is accompanied by the vulnerability dataset and source code. These have been made available as a collection of Jupyter notebooks written in Python, not only to promote transparency by empowering the independent validation of the findings contained in this report, but also to enable the cyber security community to conduct further investigations and analyses. An exhaustive exploration of the vulnerability ecosystem was prohibited by the richness of the dataset and the significant efforts in the recent literature to capture the various aspects in a streamlined and standardised form. Although the Jupyter notebooks contain a more detailed and diverse set of findings, the following ones were considered most significant and presented in the present report:

 There are significant differences between the two vulnerability measurement systems (CVSS v2 and CVSS v3), possibly attributed to the different wording of the categorical variables, fuelling subjective bias. In either case, the correlations of the three impact measures (Confidentiality, Integrity and Availability) were surprisingly low, with Integrity and Availability being less than 0.4.  There are inconsistencies and discrepancies between the different sources. Although there is an authoritative database capturing vulnerability details, this does not imply that the information in that database is accurate.  The developed taxonomies and standards used to describe the vulnerabilities are indeed rich and detailed, but only a subset of the categories were present in the 2018- 2019 vulnerabilities dataset.  There are statistically significant differences between the severity level of CVE (officially recorded) and non-CVE vulnerabilities (i.e. those that were not listed or included in the CVE databases), with the latter showing a higher score.  The exploit publication date of CRITICAL vulnerabilities is attracted near the vulnerability publication date, with most exploits being published shortly before or after the vulnerability publication date.  At least 8.65% of the vulnerabilities are exploitable. This number is expected to be higher due to zero-day exploits and the incompleteness of the datasets.  Defence Evasion, Persistence, and Discovery are the preferred tactics for the exploits.  Most exploits target web and client-side related vulnerabilities.  The top 10 weaknesses account for almost two thirds (64%) of the vulnerabilities.

STATE OF VULNERABILITIES 2018/2019

4. REFERENCES

Arbaugh, W., Fithen, W., McHugh, J. Windows of Vulnerability: A Case Study Analysis. IEEE Computer, Vol 3, No. 12, December 2000.

ENISA, Good Practice Guide for Vulnerability Disclosure: From Challenges to recommendations, 2015. Available from: https://www.enisa.europa.eu/publications/vulnerability-disclosure

ENISA, Economics of Vulnerability Disclosure, 2018. Available from: https://www.enisa.europa.eu/publications/economics-of-vulnerability-disclosure

ENISA, Threat Landscape Report, 2019. Available from: https://www.enisa.europa.eu/publications/enisa-threatlandscape-report-2018

Hutchins, E., Cloppert, M. and Amin,R. Intelligence-Driven Computer Network Defense Informed by Analysis of Adversary Campaigns and Intrusion Kill Chains. Bethesda, MD: Lockheed Martin Corporation, 2010. Available from: https://www.lockheedmartin.com/content/dam/lockheed-martin/rms/documents/cyber/LM-White-Paper-Intel-Driven- Defense.pdf

Kritikos, K., Magoutis, K., Papoutsakis, M., Ioannidis, S., A survey on vulnerability assessment tools and databases for cloud-based web application. Array, Vol 3, No. 4, 2019., pp. 1-21.

STATE OF VULNERABILITIES 2018/2019

A ANNEX: LIST OF VULNERABILITY DATABASES

A.1 FREE ACCESS DATABASES

https://exchange.xforce.ibmcloud.com/ https://www.securityfocus.com/vulnerabilities https://nvd.nist.gov/ https://www.cvedetails.com/ https://vuldb.com/ https://www.exploit-db.com/ https://www.rapid7.com/db/ https://snyk.io/features/vulnerability-database/ https://www.kb.cert.org/vuls/ https://www.first.org/global/sigs/vrdx/vdb-catalog https://help.veracode.com/reader/hHHR3gv0wYc2WbCclECf_A/lQYKhC8AvpIbz5_ULOCYMw https://www.netsparker.com/web-vulnerability-scanner/vulnerabilities/ https://www.cerias.purdue.edu/site/about/history/coast/projects/vdb.php https://wpvulndb.com https://packetstormsecurity.com/ http://cve.mitre.org/ https://0day.today/ https://www.misp-project.org/features.html https://cert.europa.eu/cert/newsletter/en/latest_SecurityBulletins_.html http://www.cnnvd.org.cn/ https://www.us-cert.gov/ics/advisories https://jvn.jp/en/ https://www.kyberturvallisuuskeskus.fi/en/homepage https://securiteam.com/ https://securitytracker.com/ https://www.zerodayinitiative.com/advisories/published/ https://www.vulnspy.com/ https://github.com/AUEB-BALab/VulinOSS https://oval.cisecurity.org/ https://seclists.org/fulldisclosure/ https://www.seebug.org/ https://cxsecurity.com/ https://en.0day.today/ https://developer.shodan.io/api/exploits/rest https://www.talosintelligence.com/ https://www.us-cert.gov/ncas/bulletins https://github.com/0x4D31/awesome-threat-detection https://www.zerodayinitiative.com/advisories/published/ https://www.zerodayinitiative.com/advisories/upcoming/

A.2 COMMERCIAL DATABASES

https://www.symantec.com/services/cyber-security-services/deepsight-intelligence https://vulndb.cyberriskanalytics.com/ https://www.flexera.com/products/operations/software-vulnerability-management.html

STATE OF VULNERABILITIES 2018/2019

https://www.accenture.com/us-en/blogs/blogs-vulnerability-intelligence https://www.auscert.org.au/services/security-bulletins/ https://www.synopsys.com/software-integrity/security-testing/software-compositionanalysis/technology/vulnerability-reporting.html https://www.cisco.com/c/en/us/td/docs/security/firepower/Application_Detectors/library-vdb/fpapp-detectors-library.html https://www.manageengine.com/vulnerability-management/help/vulnerability-databasesettings.html

STATE OF VULNERABILITIES 2018/2019

B ANNEX: JUPYTER ENVIRONMENT INFORMATION

The Jupyter notebooks used in this report can be found at: https://github.com/enisaeu/vulnreport

B.1 README.MD B.1.1 Getting started

1. Install Anaconda

Visit the Anaconda website and download the Anaconda installer for your OS (Python 3.7 version).

2. Create Environment

Clone this repository using terminal. git clone https://github.com/enisaeu/vuln-report.git Run the following from within the root of the repository conda env create --file environment.yml

3. Install required Jupyter Lab extensions

In terminal, make sure you have the enisa environment activated. conda activate enisa Once activated, install the plotly Jupyter Lab extension. jupyter labextension install @jupyterlab/plotly-extension

4. Launch Jupyter Lab

Run the following from within the root of the repository. jupyter lab If it asks you to build/rebuild make sure to accept all the prompts. You can keep an eye on the terminal window that launched Jupyter Lab to see when it's ready. You should refresh your browser once it's done.

5. Check out the example notebook

STATE OF VULNERABILITIES 2018/2019

All done, check out the example notebook located at notebooks/example.ipynb.

B.2 LIST OF JYPTER NOTEBOOKS

Jupyter Notebook Name Content

0_table_of_contents Table of contents.

Comparison of different scores

1_average_CVSS_scores

between vulnerabilities.

Top products/vendors with the most vulnerabilities, correlation

2_high_CVSS_distribution

between vendors, CVSS scores and tactics.

Top weaknesses, Top products/vendors with most 3_CWE_with_high_CVSS (unique) weaknesses, Average CVSS scores for weaknesses, CWEs with high CVSS scores.

Weaknesses regarding web

4_top_10_weaknesses_per_product

browsers and operating systems.

Vulnerabilities in different sectors,

5_top_vulnerabilities_per_sector

top CVEs in them.

CVSS scores, severity ratings, 6_vulnerabilities_in_web_vs_native exploits, vendors In web browsers and windows applications.

CVSS scores, severity ratings, 7_vulnerabilities_in_open_source exploits, vendors, weaknesses in open source projects.

8_events_attributed_to_vulnerabilities Incidents and vulnerabilities.

9_high_score_and_known_exploits Exploitation based CVSS scores.

Exploits before/after published

10_vulnerability_lifecycle

date, end of support.

ATT&CK Framework and CAPEC

11_ATT&CK_capec

patterns.

Top software categories, cvss

12_application_types

scores, ATT&CK and tactics.

Analysis of the prices of Vulndb

13_prices

and Zerodium exploit price data.

-N -EN -014 -20 -03 TP

ABOUT ENIS A

The mission of the European Union Agency for Cybersecurity (ENISA) is to achieve a high common level of cybersecurity across the Union, by actively supporting Member States, Union institutions, bodies, offices and agencies in improving cybersecurity. We contribute to policy development and implementation, support capacity building and preparedness, facilitate operational cooperation at Union level, enhance the trustworthiness of ICT products, services and processes by rolling out cybersecurity certification schemes, enable knowledge sharing, research, innovation and awareness building, whilst developing crossborder communities. Our goal is to strengthen trust in the connected economy, boost resilience of the Union’s infrastructure and services and keep our society cyber secure. More information about ENISA and its work can be found at www.enisa.europa.eu.

ISBN: 978-92-9204-325-4 DOI: 10.2824/139301